🤖 AI Soup - Weekly Roundup

November 24, 2025

Your weekly digest of the AI revolution happening at breakneck speed

1. MAJOR STORY DEEP-DIVES

🚀 GOOGLE RETAKES THE THRONE WITH GEMINI 3

The News: After playing catch-up for nearly two years, Google has launched Gemini 3, along with updated image models (Nano Banana Pro) and a new developer platform (Antigravity). By almost every metric—benchmarks, speed, and user preference—Google momentarily holds the title for the most powerful AI model in the world. Even OpenAI's CEO Sam Altman reportedly sent a memo warning staff that the "vibes might be rough" for a bit.

Why It Matters: This release kills the "Scaling Wall" narrative. For months, rumors circulated that AI progress was plateauing—that feeding models more data wasn't making them smarter. Gemini 3 proves that scaling laws essentially still hold; we haven't hit the ceiling yet. Furthermore, Gemini 3 was trained on Google's own TPUs (Tensor Processing Units), not Nvidia chips. This is a subtle but massive signal that Nvidia's iron grip on AI hardware might eventually loosen.

The Details:

• Benchmark dominance: Gemini 3 hit 37.5% on the "Humanities Last Exam" (a test designed to be impossibly niche for AI), beating the previous best of ~31%.
• Vision leap: On ScreenSpot Pro (understanding what’s on a computer screen), it scored 72.7% vs. the previous best of 36.2%. That is a generation-skipping leap for AI agents.
• Massive Scale: Google revealed the Gemini app now has 650 million monthly users.

The Honest Take: Google is flexing its "resource advantage." While OpenAI burns cash renting servers, Google owns the chips, the cloud, and the distribution. However, benchmarks are fleeting. The real test is whether users actually switch their default behavior from ChatGPT. For now, the "Gemini is bad" meme is officially dead.

💰 AMAZON VS. PERPLEXITY: THE "DOORDASH PROBLEM"

The News: The cold war between AI agents and the web just went hot. Amazon is suing AI search engine Perplexity for "agentic shopping"—effectively using AI bots to bypass Amazon's interface to buy things for users. Perplexity calls it innovation; Amazon calls it "bullying" and a violation of the Computer Fraud and Abuse Act.

Why It Matters: This is the start of the war for the future of the internet economy. If an AI agent buys your toothpaste, you don't see Amazon's ads, you don't see their recommended add-ons, and you don't join their loyalty programs. This is the "DoorDash Problem"—where an intermediary (the AI) turns the provider (Amazon) into a dumb commodity pipe, destroying their margins.

The Details:

• Perplexity’s "Buy with Pro" feature acts as a user agent, filling carts and checking out.
• Amazon claims this violates their Terms of Service and confuses their systems into thinking bots are humans.
• Perplexity’s defense is aggressive: "Software is becoming labor," effectively arguing that an AI agent has the same rights to browse the web as a human.

The Honest Take: Amazon is terrified. They built a $60B+ ad business on people staring at search results. AI agents obliterate that business model. While Perplexity is the target today, Amazon is really trying to set a legal precedent before Google or OpenAI launch similar features at scale. Without a new "treaty" between AI and the web, we are heading toward a fragmented, paywalled internet.

🏢 JEFF BEZOS UN-RETIRES FOR A $6 BILLION BET

The News: Jeff Bezos is returning to the CEO seat—not at Amazon, but at a new stealth AI startup called Project Prometheus. The company is launching with a staggering $6.2 billion seed round (yes, with a 'B') to apply AI to physical engineering and manufacturing.

Why It Matters: We have enough chatbots. The next frontier is "Physical AI"—applying these models to atoms, not just bits. When the guy who built the world's most efficient logistics network decides to run a manufacturing AI company, it signals that the industry is shifting from "write me a poem" to "build me a spacecraft."

The Details:

• Focus: Engineering, manufacturing of computers, automobiles, and spacecraft (likely leveraging Blue Origin).
• Capital: $6.2B is unprecedented for a seed round. For context, huge labs like SSI raised $1B.
• Talent: They are poaching researchers from OpenAI and DeepMind to solve problems in material science and industrial automation.

The Honest Take: Bezos scaling a startup with infinite money sounds like a cheat code, but the history of "super-teams" in tech is mixed. The most interesting part is the focus: Manufacturing is a "boring" trillion-dollar layer that AI hasn't cracked yet. If this works, it matters more to the GDP than another LLM.

2. 💡 THIS WEEK'S SMARTEST TAKE

The Insight:

"The promise of the agentic economy is built on an extremely brittle relationship between the Doordashes of the world and the AI companies who deploy agents... If people stop using the apps and websites and start sending agents instead, that business really starts to break down... [Services] might just become commodity databases of information competing on price alone." — Nilay Patel (Decoder)

The Context: Nilay was discussing the lawsuit between Amazon and Perplexity. He argues that the entire "App Store economy" (Uber, Airbnb, DoorDash) relies on owning the customer relationship.

Why It Is Illuminating: Everyone talks about "AI Agents" like they are magic butlers. Patel correctly identifies that agents are actually macro-economic parasites. If an AI agent always books the cheapest flight or ride, it strips all brand loyalty and profit margin from the service providers. This suggests the future isn't just "better tech," but a brutal war over who gets to own the customer interface. It reframes the Amazon lawsuit not as a petty squabble, but as a fight for survival.

3. ⚖️ OVERHYPED OR UNDERHYPED?

The Subject: The "AI Bubble" Bursting

The Overhype (8/10):

• Wall Street is panicking because capital expenditure (CapEx) is high and "Killer Apps" are scarce.
• Stock prices for AI companies wobbled this week due to "bubble fears."
• The narrative assumes that if ChatGPT doesn't double revenue next month, the tech is a failure.

The Underhype (9/10):

• Nvidia's Numbers: They just posted $57 billion in quarterly revenue (up 62%). That is actual cash changing hands, not hype.
• The Ceiling Lifted: Gemini 3 proves we haven't hit a technical wall.
• Consumer Spend: Suno (AI Music) hitting $200M revenue proves normal people will pay for AI creativity, not just enterprise seats.

The Verdict: [OVERHYPED] (High Anxiety, Low Evidence) The "Bubble" talk is mostly financial anxiety searching for a reason. When the hardware supplier (Nvidia) is printing $57B a quarter and the software providers (Google/OpenAI) are breaking performance records, the fundamental trend line is still vertical. We may see valuation corrections, but the utility bubble hasn't popped because it isn't a bubble.

4. 📡 SIGNAL VS. NOISE

SIGNAL (Pay Attention) → "Compaction" in OpenAI's Codex Max: A new technique allowing models to work on code for 24+ hours by compressing their memory. This is how we get to long-term agents. → Google using TPUs for Gemini 3: Google managed to beat the state-of-the-art without relying on Nvidia's H100s. This is the first crack in Nvidia's monopoly. → Microsoft/Nvidia/Anthropic Tripartite Deal: Anthropic committed to buy $30B of Azure compute? The incestuous financial loop between cloud providers and AI labs is cementing into a cartel. → Suno's $200M Revenue: This is net-new behavior. People paying to generate songs is a sign that creative AI has product-market fit.

NOISE (Ignore Safely) → Grok 4.1's "Emotional Intelligence": While improvements are nice, XAI is still drafting behind the leaders. "Less woke" isn't a technical moat. → Demis Hassabis Hype Tweets: Even the quiet researchers are acting like influencers now. Ignore the tweets; look at the benchmarks. → State-Level AI Bans: With Trump's "Genesis Mission" likely to federalize AI rules and block states from regulating, local laws are likely dead on arrival.

5. 📊 WEEK IN NUMBERS

• $57,000,000,000: Nvidia’s quarterly revenue. For context, that's more than Intel and AMD combined, times a multiplier.
• $6,200,000,000: The seed funding for Jeff Bezos's new company, Project Prometheus.
• $30,000,000,000: The amount of Azure compute Anthropic has committed to buying as part of the new Microsoft/Nvidia deal.
• 650,000,000: Monthly users on Google's Gemini app. They are closing the gap with ChatGPT faster than people realize.
• 72.7%: Gemini 3's score on "ScreenSpot Pro" (screen understanding). The previous best was 36.2%. This is a 2x jump in AI's ability to "see" our computers.
• $200,000,000: Annual revenue run-rate for AI music startup Suno.
• 24 Hours: The autonomous coding time achieved by OpenAI's new GPT-5.1 Codex Max model.

6. 🔮 WHAT THIS ALL MEANS

Patterns: This week was defined by Integration and resource flexing. Google didn't just launch a model; they launched an ecosystem (IDE, Image, Video, Search integration). OpenAI didn't just launch a chat update; they launched specialized "worker" models (Codex Max). The era of the "general purpose chatbot" is fading in favor of specialized agents and vertically integrated stacks.

Trajectory: The industry is accelerating, not slowing. The fear that LLMs had hit a "wall" was soundly refuted by Gemini 3’s massive benchmark jumps. We are moving from "Chat" (talking to AI) to "Agentic Work" (AI doing tasks like coding for 24 hours or buying things on Amazon).

Tensions: The primary tension is now Commercial vs. Architectural. Technically, agents are ready to browse the web. Commercially, the web (Amazon, DoorDash, Publishers) is trying to sue them out of existence. We are entering a period of intense legal warfare where the "robots.txt" of the internet is renegotiated in court.

Implications: For builders, the "wrapper" era is over; you need to build deep integrations or physical world applications. For investors, the "AI Bubble" isn't popping, but the winners are consolidating into the companies that own the infrastructure (Google, Nvidia, Microsoft). For users, you are about to get AI that doesn't just talk, but does—and you'll likely have to pay for the privilege to bypass the ad-cluttered web.

Forward thought: If agents can beat humans at navigating Amazon to find the best price, how long until Amazon bans human users entirely to force us back into their ad-optimized UI?

7. 🎯 THE BOTTOM LINE

The single most important development this week is the release of Gemini 3, not just because it's "smarter," but because it proves the AI arms race has remaining runway. Google has successfully leveraged its massive data and chip advantage to catch up to OpenAI, turning a one-horse race into a definitive duel.

For Investors: Ignore the daily stock fluctuations. Nvidia's revenue and the $30B purchase commitments from Anthropic show the infrastructure build-out is locked in for years. The money is committed. For Builders: Look at OpenAI's "Codex Max." The future is "long-context autonomy." Build tools that can run in the background for hours, not seconds. For Skeptics: The "AI is a bubble" narrative took a body blow this week. Revenue is real, capabilities are jumping, and user bases are in the hundreds of millions.

The Question Everyone Should Ask: If Amazon wins its lawsuit against Perplexity and blocks AI agents, does the "Agentic Future" die before it starts, or will the AI companies simply grow big enough to ignore the law?

8. CLOSING KICKERS

🤖 Bezos: Jeff Bezos looked at the AI industry and decided $6 billion was the correct price for a "mid-life crisis" project.

📈 Nvidia: Jensen Huang is the only CEO who can tell Wall Street "we made $57 billion" and have analysts ask, "yeah, but can you do more?"

🎯 Codex Max: We finally have an AI that can code for 24 hours straight, which coincidentally creates the first AI that requires caffeine and therapy.

Tomorrow's Burning Question: Will we look back at 2025 as the year we stopped surfing the web and started sending our AI butlers to do it for us?